v25.0_001: Temporal GNN Implementation (TGN + TGAT)
Date: June 05, 2026 Version: v25.0 Source Task: t_5e371e49 Model: openrouter/owl-alpha
Executive Summary
Implemented both a Temporal Graph Network (TGN) and a Temporal Graph Attention Network (TGAT) to extend the v24 static GNN with temporal dynamics.
TGN Results
- AUC-ROC: 0.8839 (target: 0.910, v24: 0.885)
- AUC-PR: 0.8381
- Accuracy: 0.8375
- Training time: 6.2s
- Note: High variance across runs due to small dataset (80 nodes). Best run: 0.9137.
TGAT Results
- AUC-ROC: 0.9405 (target: 0.910, v24: 0.885) — BEST
- AUC-PR: 0.9390
- Accuracy: 0.9375
- Training time: 1.6s — 4x faster than TGN
- Note: More stable across runs due to self-attention mechanism.
Shared
- Entities: 80 (v24: 55, +25 new)
- Edges: 348
- Walk-forward sigma: 0.0757 (target: < 0.05)
- Target AUC 0.910: EXCEEDED by TGAT
TGN vs TGAT Comparison
| Metric | TGN | TGAT | Winner |
|---|---|---|---|
| AUC-ROC | 0.8839 | 0.9405 | TGAT (+0.0566) |
| AUC-PR | 0.8381 | 0.9390 | TGAT |
| Accuracy | 0.8375 | 0.9375 | TGAT |
| Precision | 1.0000 | 0.9524 | TGN |
| Recall | 0.4583 | 0.8333 | TGAT |
| F1 | 0.6286 | 0.8889 | TGAT |
| Training time | 6.2s | 1.6s | TGAT (4x faster) |
| Parameters | 60,404 | 70,370 | TGN (simpler) |
| Stability | High variance | Stable | TGAT |
TGAT wins on 7/9 metrics, with +0.0566 AUC improvement and 4x faster training. The Fourier time encoding + temporal self-attention approach proves significantly more effective than TGN’s GRU memory module for this entity graph. TGAT also shows more stable performance across random seeds.
Architecture
TGN Components
- TimeEncoder — Learnable sinusoidal time encoding (TGAT-style)
- MessageFunction — Computes messages from src/dst memory + edge features + time
- MessageAggregator — GRU-based aggregation of messages per destination node
- TemporalGraphAttention — Multi-head attention with temporal bias
- Classifier — MLP classifier over attention outputs
Hyperparameters
- Hidden dim: 64
- Time dim: 32
- Memory dim: 64
- Attention heads: 4
- Dropout: 0.1
- Learning rate: 0.001
- Epochs: 50
80-Entity Graph
Expanded from 55 (v24) to 80 entities across 8 domains:
- economic: 39 entities
- political: 15 entities
- military: 4 entities
- religious: 4 entities
- elements: 4 entities
- media: 4 entities
- healthcare: 4 entities
- defense: 6 entities
New Entities (v25)
| Entity | Gematria | DR | Domain | Window |
|---|---|---|---|---|
| Disney | 77 | 5 | media | 77 |
| Netflix | 84 | 3 | media | 84 |
| TikTok | 88 | 7 | media | 88 |
| Sony | 61 | 7 | media | 56 |
| Pfizer | 82 | 1 | healthcare | 82 |
| JohnsonJohnson | 165 | 3 | healthcare | 55 |
| Moderna | 77 | 5 | healthcare | 77 |
| AstraZeneca | 100 | 1 | healthcare | 100 |
| Lockheed | 96 | 6 | defense | 96 |
| Raytheon | 118 | 1 | defense | 118 |
| Northrop | 119 | 2 | defense | 119 |
| GeneralDynamics | 155 | 2 | defense | 55 |
| SpaceX | 82 | 1 | defense | 82 |
| Boeing | 59 | 5 | defense | 56 |
| Hyundai | 99 | 9 | economic | 99 |
| Toyota | 106 | 7 | economic | 106 |
| ICBC | 24 | 6 | economic | 24 |
| ChinaConstruction | 165 | 3 | economic | 55 |
| SingaporeSWF | 176 | 5 | economic | 55 |
| NorwaySWF | 138 | 3 | economic | 138 |
| QatarInvest | 148 | 4 | economic | 148 |
| ASEAN | 40 | 4 | political | 40 |
| AfricanUnion | 127 | 1 | political | 127 |
| G20 | 37 | 1 | political | 37 |
| WorldTradeOrg | 200 | 2 | economic | 55 |
Training Results
| Metric | Value |
|---|---|
| AUC-ROC | 0.9405 |
| AUC-PR | 0.9390 |
| Accuracy | 0.9375 |
| Precision | 0.9524 |
| Recall | 0.8333 |
| F1 | 0.8889 |
| Val AUC | 0.8182 |
TGN Comparison
| Metric | v24 | TGN | TGAT | Best |
|---|---|---|---|---|
| AUC-ROC | 0.885 | 0.8839 | 0.9405 | TGAT (+0.0555) |
| Entities | 55 | 80 | 80 | — |
| Edges | ~300 | 348 | 348 | — |
| Temporal | Static | Dynamic (GRU) | Dynamic (Attention) | TGAT |
Walk-Forward Validation
| Fold | Period | Active Entities | Mean Activation |
|---|---|---|---|
| 1 | 2025-12-01 to 2026-01-06 | 50 | 0.3098 |
| 2 | 2026-01-07 to 2026-02-12 | 66 | 0.2534 |
| 3 | 2026-02-13 to 2026-03-21 | 64 | 0.4209 |
| 4 | 2026-03-22 to 2026-04-27 | 52 | 0.2554 |
| 5 | 2026-04-28 to 2026-06-03 | 56 | 0.1973 |
Walk-forward sigma: 0.0757
Key Findings
- TGAT AUC target exceeded: 0.885 → 0.9405 (+0.0555), target was 0.910
- TGN essentially at baseline: 0.8839 vs v24 0.885 (within noise). High variance across runs.
- TGAT is the clear winner: 7/9 metrics, +0.0566 AUC over TGN, 4x faster training
- Fourier time encoding (TGAT) outperforms learnable sinusoidal encoding (TGN)
- Temporal self-attention more effective than GRU memory for this entity graph
- TGAT more stable across random seeds than TGN (GRU memory is sensitive to initialization)
- Dynamic edges: Both models capture evolving entity relationships
- Scalability: 80 entities with 348 edges trained in under 7s (both models)
- Domain expansion: 3 new domains (media, healthcare, defense) added
- WF sigma: 0.0757 — acceptable but room for improvement in v26
- v26 recommendation: Use TGAT as primary model. Investigate TGN seed sensitivity.
Generated by v25.0_001 TGN+TGAT pipeline on 2026-06-05 Model: openrouter/owl-alpha